Possible salt and pepper distributions

On a set of $(10000,784)$ $0$-matrices renoising with probability $p$ is applied. For different noise distributions.

Actually used

$a \in \{-1,0,1\}$ chosen with equal probability. (As matrix of img.shape dimensions)
$b \in \{0,1\}$ chosen with probability $p$. (As matrix of img.shape dimension)
$img=img+10 \cdot a \cdot b$ and then clipped to $[-1,1]$

For low $p$ the image is not affected and for $p=1$ half of the pixels are set to either $-1$ or $1$.
This was the noise/renoising used until now.

Modify all

$a \in \{-1,1\}$ chosen with probability $p$. (As matrix of img.shape dimension)
$img=img+10 \cdot a$ and then clipped to $[-1,1]$

For low $p$ theimage is all set to $-1$ and for $p=1$ all pixels are set to either $1$. But in every steps all pixels are affected, i.e. no pixel left at $0$.
To destructive the underlying information is destroyed

Balanced modify all

$a \in {0,1}$ chosen with probability $p$. (As matrix of img.shape dimension)
$b \in \{-1,1\}$ chosen with equal probability. (As matrix of img.shape dimensions)

$img=img+10 \cdot a \cdot b$ and then clipped to $[-1,1]$

For low $p$ theimage is not affected and for $p=1$ all of the pixels are set to either $-1$ or $1$. But in each step not all pixels are modified but only a $p$ percent of them

New proposed noise/renosing method.

[NbConvertApp] Converting notebook Noise_dist.ipynb to html
[NbConvertApp] Writing 5969339 bytes to ../../Page/Iterative-autoencoder-networks-for-signal-denoising/Noise_dist.html